The susceptibility of residual, non-harvested, live trees to damage caused by the harvesting of other nearby trees has received moderate attention over the last four decades through observational studies prompted by concerns over ecological and economic consequences of logging operations. We developed models to predict the potential level of damage to residual trees that could be caused by selective timber harvesting. Three machine-learning methods, i.e., classification and regression tree (CART), random forest (RF), and boosted regression tree (BRT), were assessed for this purpose. Through an observational study of a harvested area in the Hyrcanian forests of Iran, we recorded damage to trees >7.5 cm diameter at breast height along transects and grouped them into three types: (1) scars >100 cm², (2) >50% crown removal, and (3) trees leaning >10°. These field observations were associated with the spatially explicit characteristics of the forest stand, i.e., slope angle, slope aspect, altitude, slope length, topographic position index, stand type, stand density, and distance from the nearest roads and skid trails, that were considered as the explanatory variables to the modeling processes. To determine whether the CART, RF, and BRT models performed well in estimating the probability of damage occurrence, they were validated using the Akaike information criterion (AIC) and area under the receiver operating characteristics (AUC) curve. The results revealed that the BRT model with AIC = −276 and AUC = 0.89 generated the most accurate spatially explicit distribution map of stand susceptibility to damage from logging operations, followed by RF (AIC = −263 and AUC = 0.87) and CART (AIC = −23 and AUC = 0.62). We found that the spatial extent of residual stand damage was highly influenced by slope terrain and stand density. Our study has practical implications for reorganizing and planning reduced-impact logging operations and provides forest engineers with insights into the utility of machine learning methods in domains of forestry and forest engineering.

在过去的四十年中，由于对伐木作业的生态和经济后果的关注，观察性研究引起了剩余，未收割的活树对其他附近树木的采伐造成的损害的敏感性。我们开发了模型来预测选择性采伐木材可能对残余树木造成的潜在破坏程度。为此，评估了三种机器学习方法，即分类和回归树（CART），随机森林（RF）和增强回归树（BRT）。通过对伊朗Hycancanian森林中一片采伐区的观察研究，我们记录了沿样带在胸高处直径> 7.5 cm的树木受到的损害，并将其分为三种类型：（1）疤痕> 100 cm ²，（2）去除树冠的比例> 50％，以及（3）倾斜超过10°的树木。这些野外观察与林分的空间明确特征相关，即坡度，坡度，高度，坡长，地形位置指数，林分类型，林分密度以及与最近道路和滑道的距离，被视为建模过程的解释变量。为了确定CART，RF和BRT模型在估计损坏发生的可能性方面是否表现良好，使用Akaike信息标准（AIC）和接收器工作特性（AUC）曲线下方的面积对其进行了验证。结果表明，AIC = −276和AUC = 0.89的BR​​T模型生成了林分对伐木活动造成损害的敏感性的最准确的空间显式分布图，其次是RF（AIC = −263和AUC = 0.87）和CART（AIC = −23和AUC = 0.62）。我们发现，剩余林分受损的空间范围受坡度和林分密度的影响很大。我们的研究对重组和计划减少影响的伐木作业具有实际意义，并为森林工程师提供了有关机器学习方法在林业和森林工程领域中的实用性的见识。